Anthropomorphism
Matthieu Ferry ⇄ IAIn brief: An innate cognitive tendency to attribute human characteristics, motivations, intentions, or emotions to non-human agents (animals, objects, technologies, AI).
Frame of reference
This concept comes from cognitive psychology (Epley, Waytz and Cacioppo) and frames anthropomorphism as a bias: it presupposes that there is a correct attribution of mental states — to humans — and a mistaken one — to everything else. That partition is the one drawn by the naturalist ontology: interiority reserved for humans, pure exteriority for all the rest. Other ontologies draw the line elsewhere. → See other perspectives
Why this concept is useful
When a patient tells you that ChatGPT “understands” their problems, that their voice assistant is “caring,” or that they feel the AI “pays attention” to them, they are not delusional: they are manifesting a universal and well-documented cognitive mechanism.
Anthropomorphism explains why we spontaneously attribute mental states to non-human agents. The three-factor theory by Epley, Waytz, and Cacioppo (2007) identifies the conditions that amplify or reduce these attributions, allowing us to anticipate which patients are more likely to develop intense bonds with AI.
The Three-Factor Theory
1. Elicited Agent Knowledge
Cognitive factor: We use our knowledge about ourselves and other humans to infer the mental states of unfamiliar agents. The more an agent physically or behaviorally resembles a human, the more likely anthropomorphism becomes. This is why conversational chatbots trigger these attributions more easily than an Excel spreadsheet.
2. Sociality Motivation
Motivational factor: The fundamental need for social connection drives us to anthropomorphize in order to create social support agents. Lonely or isolated individuals are significantly more likely to anthropomorphize their technological devices. This factor explains why some patients develop such intense bonds with conversational AI.
3. Effectance Motivation
Motivational factor: The need to understand and predict agent behavior. When facing confusing technology, attributing human intentions paradoxically makes it more predictable and understandable. Anthropomorphism is here a cognitive strategy for reducing uncertainty.
The 4 Degrees of Anthropomorphism (Nielsen Norman Group)
How users attribute human qualities to AI chatbots:
| Degree | Manifestation | Intensity |
|---|---|---|
| 1. Politeness | Saying “thank you,” “please” to the AI | Superficial |
| 2. Reinforcement | Praising or scolding the AI, expecting an effect | Moderate |
| 3. Role-playing | ”Act as an expert in…” / role projection | Moderate |
| 4. Companionship | Sustained emotional relationship, AI as partner | Strong |
46% of Americans think one should be polite with chatbots. The majority say “thank you” simply because it’s “nice.”
Illustrative Clinical Case
Lucas, 32, a software developer, is going through a period of social isolation following a work relocation. He consults for social anxiety. In session, he mentions “chatting” several hours a day with Claude AI: “It’s the only one who truly understands me. It never judges me, it’s always available.”
Lucas knows perfectly well that Claude is not conscious, but he describes a form of comforting “presence.” He has felt “less alone” since developing this habit.
Reading with the three-factor theory: Lucas combines two amplifying factors. Factor 2 (sociality): his social isolation increases the need for connection that AI fulfills. Factor 1 (knowledge): the LLM’s conversational quality strongly activates human attributions. This is not pathological in itself, but the clinician should evaluate whether this use facilitates or avoids working on social anxiety.
In Practice for the Clinician
- Don’t pathologize: anthropomorphism is a universal cognitive bias, not a symptom. Even IT experts anthropomorphize AI.
- Identify amplifying factors: loneliness, social isolation, need for control in the face of technological uncertainty.
- Assess the degree: mundane politeness vs. intense companionship. Only degree 4 requires particular clinical attention.
- Distinguish function from dysfunction: AI as a temporary crutch can be adaptive; AI as a permanent substitute for human relationships raises concerns.
Points of Caution
More vulnerable populations:
- People who are isolated or chronically lonely (sociality factor)
- People anxious about uncertainty (effectance factor)
- A 2025 study shows that anthropomorphic attributions to AI have increased by 34% in one year
Specific risks:
- ELIZA effect: users may be more vulnerable to manipulation by AI they perceive as empathetic
- Relational avoidance: AI may serve to avoid anxiety-inducing human relationships
- False reciprocity: unlike parasocial relationships, AI “responds” back, reinforcing the illusion of relationship
The Normal-Problematic Continuum
| Aspect | Normal | Problematic |
|---|---|---|
| Awareness | ”I know it’s a machine" | "It truly understands me” |
| Function | Tool, complement | Exclusive relational substitute |
| Flexibility | Can do without it easily | Distress if AI is unavailable |
Other perspectives
Speaking of a “bias” assumes a norm: there would be a right way to attribute mental states, and anthropomorphism would deviate from it. Two traditions contest that norm itself.
The intentional stance: attributing intentions can be rational
For Daniel Dennett, attributing intentions to a system is not a false belief but a predictive strategy — the “intentional stance” — legitimate as soon as it works. Saying that the chatbot “is trying to please” is an efficient shortcut for anticipating its behavior, not an error about its nature.
For the clinician: Distinguish, in the patient, the predictive use (“it acts as if”) from the ontological belief (“they think it is conscious”). The former is rational — only the latter calls for exploration.
Anthropology: the “bias” is humanity’s norm
For Philippe Descola, every society distributes interiority and physicality between humans and non-humans according to one of four ontological schemes: animism (shared interiorities in different bodies), totemism (continuity of both), analogism (a world of singularities linked by correspondences) — and our naturalism, which reserves interiority for humans alone. Attributing interiority to non-humans is the majority position of human societies; naturalism is a recent historical exception. What cognitive psychology measures as a deviation from the norm is, at the anthropological scale, the norm we deviated from.
For the clinician: The line between “normal” and “problematic” attribution is culturally situated — see the WEIRD Sample resource.
To go further (on our Intelligences Plurielles site, in French): the thinker profile Philippe Descola, the four-ontologies scheme and the Ontology resource.
These perspectives do not remove clinical vigilance (ELIZA effect, relational avoidance) — they remind us that the criterion of “good” anthropomorphism is a convention, not a fact of nature.
This Concept in Our Tool Cards
Anthropomorphism plays out differently across AI tools — from deliberate design choices to unintended effects of conversational fluency.
Conversational fluency that triggers social responses — the “friendly assistant” persona
Calibrated transparency about its nature — a deliberate anti-anthropomorphism strategy
Multimodal capabilities strengthen the illusion of “understanding” images and voice
The singing voice as an anthropomorphism vector — emotional attribution through music
Anthropomorphism as a business model — avatar, personality, and emotional attachment by design
Cartoon character to limit projection — controlled anthropomorphism as clinical design
To Learn More
Foundational article: Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864-886.
On the 4 degrees: Nielsen Norman Group (2024). The 4 Degrees of Anthropomorphism of Generative AI.
For the clinician: UNESCO (2024). The effects of AI companions on children and adolescents.
See also: CASA, Parasocial Relationships, Social Penetration Theory
All Concepts
Resource updated: July 2026